Research map / Label-efficient and robust learning / Recognition and long-tail
Effective and Efficient Representation Learning for Flight Trajectories
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Flight2Vec, a flight-specific trajectory representation learning method using behavior-adaptive patching and a 26-class moving-direction loss, evaluated on flight trajectory prediction, recognition, and anomaly detection.
Problem
Learning a unified representation for flight trajectories is hard because behaviors (turns, holding, takeoff/landing) are sparse relative to straight-line cruise and because 3D spatial continuity is poorly captured by MSE-based reconstruction.
Approach
A behavior-adaptive patching scheme that builds patches around detected behavior (angle-change) points and down-samples non-behavior segments, combined with a masked-patch pretraining objective that adds a 26-direction moving-trend classification loss to MSE.
Key results
- On FTP (SCAT), Flight2Vec+BE achieves the best MAE/RMSE/MDE across all horizons, e.g., MDE 1.44 at horizon 15 versus FlightBERT++ 1.99.
- On anomaly detection (SCAT), Flight2Vec reaches AUC 0.9050 on Go-Around versus 0.8200 for DDM.
- On flight recognition (ATFMTraj), Flight2Vec is comparable to ATSCC (e.g., ACC 0.9947 vs 0.9946 on RKSla) while being about 10x faster (3.11 ms vs 43.89 ms).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add motion direction loss (ablation) | Flight2Vec · SCAT FTP, horizon 15 | RMSE altitude: 23.11 -> 18.24 | Helped |
| Add behavior-based patching instead of fixed-interval random sampling (ablation) | Flight2Vec · SCAT FTP, horizon 15 | RMSE altitude: 19.12 -> 18.24 | Helped |
| Add BE features | Flight2Vec · SCAT FTP, horizon 1 | MDE: 0.32 -> 0.29 | Helped |
| Use Flight2Vec rather than DDM | SCAT anomaly detection, Go-Around · SCAT anomaly detection, Go-Around | AUC: 0.8200 -> 0.9050 | Helped |
| Use Flight2Vec rather than ATSCC | ATFMTraj flight recognition, ESSA · ATFMTraj flight recognition, ESSA | ACC: 0.9990 -> 0.9990 | No clear effect |
| Use Flight2Vec rather than PatchTST | SCAT FTP · SCAT FTP | MTC: 2.20 -> 3.12 | Hurt |
| Use Flight2Vec rather than DDM | SCAT AD · SCAT AD | MTC: 6.53 -> 3.11 | Helped |
Surprising findings
- FlightBERT++ outperforms Flight2Vec on short-horizon predictions (horizons 1 and 3) without the BE feature fusion, which the authors attribute to attribute fusion.
Limitations
- Anomaly detection uses synthetically injected anomalies because real labeled anomalies are unavailable.
- Flight recognition results are comparable to, not better than, the ATSCC baseline.
- FlightBERT++ is better for short-horizon prediction without feature fusion.
Experimental setup
Datasets: SCAT ATFMTraj
Benchmarks: SCAT ATFMTraj
Metrics: MAE MAPE RMSE MDE ACC PRE REC AUC AUPR MTC
Baselines: LSTM+Attention PatchTST FlightBERT++ SPIRAL ATSCC DMDN DDM
Builds on: PatchTST FlightBERT++ masked autoencoding
Closest papers by meaning
- GTR: A General, Multi-View, and Dynamic Framework for Trajectory Representation Learning (ICML 2025 · similarity 0.86)
- FlightBERT++: A Non-autoregressive Multi-Horizon Flight Trajectory Prediction Framework (AAAI 2024 · similarity 0.84)
- Towards Generalizable Trajectory Prediction using Dual-Level Representation Learning and Adaptive Prompting (CVPR 2025 · similarity 0.84)
- T4P: Test-Time Training of Trajectory Prediction via Masked Autoencoder and Actor-specific Token Memory (CVPR 2024 · similarity 0.84)
- TrajMamba: An Efficient and Semantic-rich Vehicle Trajectory Pre-training Model (NeurIPS 2025 · similarity 0.83)
- Aeolus: A Multi-structural Flight Delay Dataset (NeurIPS 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.